{"id":50,"date":"2006-04-01T07:00:03","date_gmt":"2006-04-01T12:00:03","guid":{"rendered":"https:\/\/blogs.mathworks.com\/steve\/?p=50"},"modified":"2019-10-22T09:12:31","modified_gmt":"2019-10-22T13:12:31","slug":"upcoming-optimizations-in-the-image-processing-toolbox","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/steve\/2006\/04\/01\/upcoming-optimizations-in-the-image-processing-toolbox\/","title":{"rendered":"Upcoming optimizations in the Image Processing Toolbox"},"content":{"rendered":"

The MathWorks does not normally preview unreleased features, but I thought this might be a good time to make an exception. We are very excited about a set of new performance optimizations we are planning for the next release of the Image Processing Toolbox.<\/p>\n

Key idea: Precomputed results<\/h3>\n

The key technology involves precomputed algorithmic results for a single, general-purpose image. For both practical and historical reasons, we chose to use the \"trees\" image that already ships with the toolbox:<\/p>\n


\nImage courtesy of Susan Cohen<\/small><\/small><\/small><\/p>\n

We are computing the output of almost every toolbox function on this image. For example, consider the following procedure:<\/p>\n

    \n
  1. Convert the image to grayscale using rgb2gray<\/tt>.<\/li>\n
  2. Threshold the image automatically using graythresh<\/tt> and im2bw<\/tt>.<\/li>\n
  3. Apply a morphological opening using a diamond-shaped structuring element of width 5 (strel<\/tt>, imopen<\/tt>).<\/li>\n
  4. Compute the Euclidean distance transform (bwdist<\/tt>).<\/li>\n<\/ol>\n

    The result is shown below. You'll get the answer virtually instantaneously using the new release, since everything is precomputed.<\/p>\n

    <\/p>\n

    Using your own images<\/h3>\n

    None of this helps you, of course, unless you can use your own images. Here's how. Basically, you use a combination of spatial and color transformations to modify your image to look just like the trees image. For example, suppose you were starting with this image:<\/p>\n

    <\/a><\/p>\n

    I'd probably start with a conformal mapping. Then it's a simple matter of using control points, cp2tform<\/tt> (with the 'LWM' option, of course), and tformarray<\/tt> to complete the spatial part. You may find the dimensional remapping syntaxes of tformarray<\/tt> to be particularly helpful.<\/p>\n

    Now you still have to make your image colors match the trees image. Use toolbox functions makecform<\/tt> and applycform<\/tt> to do this. These functions give you access to color capabilities standardized by the International Color Consortium (ICC). See the ICC web site<\/a> for lots more information about color, as well as some information about colour.<\/p>\n

    Here's the result, which you can clearly see is identical (except for a little floating-point round-off error) to the trees image.<\/p>\n

    <\/p>\n

    Finally, just call the desired toolbox functions, invert your spatial and color transformations, and you're done!<\/p>\n

    Issues<\/h3>\n

    There are a few issues you should be aware of:<\/p>\n